Listening to Your Heart (Dunn et al. 2010)
The study three wiki pages have been asking for by name. does-somatic-feedback-guide-decisions lists it under “what would settle it.” barnaby-dunn says the wiki “should not state his 2010 position until that ingest happens.” iowa-gambling-task records that interoception “is never measured” and points here. It is now read.
The short version: Dunn et al. (2006) ended by proposing that someone look at “how variance in interoceptive ability relates to IGT performance.” Four years later the same first author did it — and the answer is that it does not relate. It moderates.
The result, and why the nulls are the finding
Read the correlations first, because the abstract does not lead with them:
| relationship | Study | result |
|---|---|---|
| interoceptive accuracy ↔ arousal ratings | 1 | ns |
| interoceptive accuracy ↔ valence ratings | 1 | ns |
| interoceptive accuracy ↔ overall HR response | 1 | ns |
| interoceptive accuracy ↔ intuitive ability | 2 | ns (r = .08, p = .46) |
| interoceptive accuracy ↔ bodily differentiation | 2 | ns (r = .07, p = .56) |
| HR response ↔ arousal ratings | 1 | r = .38, p < .01 |
| bodily differentiation ↔ intuitive ability | 2 | r = .41, p < .001 |
| HR × interoception → arousal | 1 | ΔR² = .12, p < .01 |
| HR × interoception → valence | 1 | ΔR² = .00, p = .98 |
| bodily differentiation × interoception → intuitive ability | 2 | ΔR² = .08, p < .01 |
Interoceptive accuracy, measured on its own, predicts nothing in this paper. It has no relationship to how aroused people feel, how pleasant they feel, how their hearts respond, how well they decide, or how informative their bodies are. Everything it does, it does in a product term.
That is the honest statement of the finding, and it is more interesting than “interoception shapes emotion experience.” Interoception is not a quantity of goodness. It is a coupling coefficient.
The null is contested — added by a later ingest
This page states the nulls as the finding, and they are. But Pollatos, Kirsch & Schandry (2005), read after this page was written, ran nearly Study 1 five years earlier and found the arousal main effect: good heartbeat perceivers rated affective pictures as more arousing, F(1, 39) = 5.90, P < .05, r = 0.34 with the heartbeat score. Same paradigm — IAPS at 6 s, SAM arousal and valence, Schandry counting, healthy adults, n = 44 against 58.
Two differences carry the load, and neither paper can adjudicate them:
- Pollatos et al. used extreme groups (22 good perceivers selected from a screen of ~140, plus matched comparisons). More power for a group difference than an unselected sample, and an inflated correlation across it — so their r = 0.34 and this paper’s r = .08 are not rival estimates of one quantity.
- Pollatos et al. never measured the body. ECG was recorded and used only to score the counting task, so no bodily response to the pictures exists in their dataset. They cannot test the model this paper argues for; they measured only its margin.
That second point cuts both ways, and the awkward direction is worth stating on this page rather than only on theirs. A main effect is arguably what this paper’s own interaction predicts on the margin. If arousal is b1*(HR) + b3*(HR × accuracy) with no accuracy main effect, then averaging over HR responses that are not centred on zero — and every image here produced HR deceleration — leaves an accuracy slope of b3 × (mean HR response). Pollatos et al. measured exactly that margin and found exactly that slope. The account is incomplete, because this paper’s own marginal was null and the reading owes an explanation of why; sampling supplies a candidate. Recorded as the wiki’s arithmetic, not either paper’s claim.
What the disagreement does not touch: the valence null, which both papers report (Pollatos F(1, 39) = 0.14; here delta-R-squared = .00). See core-affect.
A third reading, added with the Van der Does et al. (2000) ingest, and it is unkind to this paper’s nulls. The wiki’s account above discounts Pollatos’s selection as inflationary. Van der Does et al. — pooling 709 participants across seven studies — report that only about 8% of healthy adults meet an accuracy criterion while over 95% produce a count, and argue the task is valid for that minority and measures something else in everyone else. If that is right, Pollatos’s screen of ~140 down to 22 good perceivers (≈16%) is approximately isolating the subpopulation the task works on, and this paper’s unselected samples of 58 and 92 are computing correlations across a mixture in which roughly four in five participants contribute noise on the moderator. Mixtures attenuate toward zero. r = .08 and r = .07 are the shape that predicts.
That does not rescue the arousal main effect, and it does not touch the interaction — a diluted moderator should attenuate the interaction too, so if anything the mixture reading makes ΔR² = .12 and .08 harder to explain rather than easier, which is a point in this paper’s favour that the wiki should not skip past. What it does is put a third live explanation on the table (sampling inflation, missing bodily measurement, mixture attenuation), all of which predict the observed pattern and none of which this literature separates. This is the wiki’s inference; Van der Does et al. wrote a decade earlier about panic disorder and discuss neither study. See is-the-heartbeat-counting-task-valid.
And one further problem for the moderator specifically. This paper treats interoceptive accuracy as a stable individual difference — measured once, after the main task, and entered as a trait. Van der Does et al. report that fewer than half of accurate panic patients were still accurate at a second session, and that exercise manufactures accurate perceivers above ~100 bpm who vanish by ~95. A moderator that is partly state and partly manufacturable by arousal, in a study where the moderated variable is arousal-linked cardiac response, is a confound this design does not address. See heartbeat-detection-task.
What it should change in how this page is read: not the moderation result, which Pollatos et al. do not test, but the confidence with which the wiki repeats “interoceptive accuracy predicts nothing.” That is this study’s finding, in this study’s sample. The pre-2010 literature it argues against — Schandry (1981), Wiens et al. (2000), Critchley et al. (2004), Ferguson & Katkin (1996), Montoya et al. (1993) — mostly found otherwise, and the wiki had taken this paper’s word for that literature without reading any of it. See is-more-interoceptive-awareness-better.
What the moderation argument buys, which no main-effect study could
Worth stating precisely, because it is the paper’s real methodological contribution and the reason it belongs on does-somatic-feedback-guide-decisions rather than only in the decision-making literature.
The standing critic’s position (Moors 2009, quoted in the introduction) is that bodily responses arrive late and are therefore a consequence of cognitive-affective activity rather than a cause. That position makes a prediction nobody had tested: if the body is downstream, then how well someone perceives their body should be irrelevant to how tightly bodily change couples to feeling and choice. Perception of an epiphenomenon cannot feed back into what produced it.
Interoceptive accuracy moves the coupling in both studies. So either the body is doing work, or interoceptive accuracy is correlated with something else that determines both bodily responsivity and its relation to report — and the second is not idle, given the cardiodynamic confound (a louder heart is both easier to count and a larger HR signal). The authors do not consider that alternative. It is the main thing standing between this design and the inference it wants.
Study 1: the body grounds arousal, not valence
Fifty-eight participants viewed 25 IAPS-style images (5 each positive, neutral, sad, disgusting, fearful) for 6 s while ECG ran, rating felt arousal and felt valence for each. Interoception was measured afterward by heartbeat counting over six intervals (2 × 25 s, 2 × 35 s, 2 × 40 s).
All images produced HR deceleration — the standard orienting response (Bradley 2000) — and deceleration was greater for negative images. The correlational result is that less deceleration went with higher felt arousal (r = .38). The authors’ reading is that high-arousal individuals move faster through the “defense cascade,” so their initial orienting deceleration is smaller and their action-preparation acceleration arrives sooner. That is post hoc, and offered as such.
The moderation is the point:
- Arousal: HR × interoception, ΔR² = .12, p < .01. The better people could count their heartbeats, the more tightly their cardiac response tracked how aroused they said they felt.
- Valence: ΔR² = .00, p = .98. Nothing. Not attenuated — absent.
- The difference between them is itself significant, Z = 2.45, p < .01.
That last line matters, and the wiki should credit it: a dissociation claimed from one significant and one non-significant effect is the most common statistical error in this literature, and Dunn et al. did not make it. They tested the difference.
What it means for core-affect. Core affect’s most economical claim is that feeling is valence plus arousal. This paper does not touch the claim directly, but it separates the two dimensions by their grounding: arousal is bodily and gated by bodily perception, valence is not. Combined with Ferré et al.’s (2024) finding that “feeling” ratings are near-orthogonal to valence norms across 1,286 words, the wiki now has two independent sources — one physiological, one lexical, fourteen years apart — saying that the felt/bodily dimension of emotion and the hedonic dimension come apart. Neither refutes core affect. Both make the two-dimensional package look less like one thing.
And it is a point for the two-factor lineage. Strong Jamesian accounts predict interoception should gate valence experience too; hybrid accounts (Schachter & Singer 1962; Frijda 2007, who holds that valence is “not bodily at all” but a blend of appraisal and action tendency) predict arousal only. The data fit the hybrid. Given that Dror (2017) shows the two-factor theory did not arise from the James–Lange debate at all, there is a small irony worth recording: the theory that reached the right prediction here got there from social-influence research, not from thinking about bodies.
Study 2: the mixed blessing
Ninety-two participants played the intuitive-reasoning-task — 100 trials, four decks, two profitable (60% win) and two unprofitable (40% win), crossed with reinforcement magnitude — while EDA and HR were recorded in the pre-decision window. Interoception was measured afterward, as in Study 1.
The somatic marker finding replicates, in a cleaner task. Anticipatory bodily responses differentiated profitable from unprofitable decks, and the composite bodily-differentiation index correlated with intuitive ability at r = .41. Bodies that told the truth belonged to people who chose well.
And the 2 × 2 answers Tomb et al. (2002). Bodily responses tracked profitability and were completely unrelated to magnitude — no main effect, no interaction, Fs < 1. This is more consequential than the paper makes of it. skin-conductance-response and iowa-gambling-task both carry the variance interpretation as the sharpest live objection to the anticipatory SCR: in the original IGT the bad decks were also the high-variance decks, so “the body marks badness” and “the body tracks variance” were confounded from the outset. Dunn et al. cross the two factors and the body follows goodness, not variance. The confound is broken, in a design built by the person who catalogued it.
Hold two qualifications. This is a different task — it does not retrospectively decontaminate IGT data, where the confound remains exactly as Tomb et al. described. And the EDA main effect that half the composite rests on is p = .05, reported as “a strong trend toward a significantly greater” response, which at the paper’s own alpha it is not.
Then the interaction, which is what the paper is for. Bodily differentiation × interoceptive accuracy predicted intuitive ability, ΔR² = .08. Deconstructed: when the body favoured the good decks, better interoception went with better decisions; when the body favoured the bad decks, better interoception went with worse ones.
So the answer to does-somatic-feedback-guide-decisions’s question 3 — does perceiving bodily signals help? — is: it depends entirely on whether your body is right. Interoception is a gain term. It amplifies signal and noise alike, and nothing in the perceiving distinguishes them.
The figure problem nobody tested
The single thing this wiki should hold most sceptically, and it is not mentioned in the paper.
Both key figures plot three regression lines at −1 SD, mean, and +1 SD of interoceptive accuracy, per Aiken & West (1991). In Figure 1a and Figure 5 alike, the worse-interoception line slopes negative — poor perceivers show less felt arousal with less HR deceleration, and better intuitive ability when their bodies favour the bad decks.
If that is real, it is a crossover interaction, and no version of any bodily-feedback theory predicts it. Poor perceivers should be uncoupled from their bodies, not inversely coupled to them.
If it is not real, it is arithmetic: the plotted simple slope at −1 SD is b₁ − b₃, and when the interaction term is comparable in size to the main effect — which here it is — a slightly negative line falls out of the model whether or not anything in the data reverses.
The paper cannot tell you which, because it never tests the simple slopes against zero, though the procedure it cites includes doing so. It reports only that coupling grew “stronger” as interoception increased, which describes magnitude and silently passes over sign. The distinction is not cosmetic: “the body matters only for people who can feel it” and “the body misleads people who cannot feel it” are different theories with different clinical implications, and this design was capable of separating them.
What this does to the wiki’s somatic-marker thread
Read in order — bechara-2005-somatic-markers → dunn-2006-somatic-marker-evaluation → this — the sequence is one researcher reading a framework, cataloguing what its evidence cannot support, naming the study nobody had run, building a task that fixes the flaws he documented, and running it. That is rare enough to note. It is also why this paper deserves more sceptical attention than a friendly reading gives it: the person who designed the test had a published position on what the test should find.
The verdict it returns is not the one either side wanted:
- For the framework: the somatic signal is real, tracks advantage, and predicts choice (r = .41) in a task purged of the reversal confound and the variance confound. That is better evidence for somatic marking than the iowa-gambling-task ever supplied, and it arrives from the critics.
- Against the framework: the signal is not sufficient, not necessary, and not the whole story — 27% of a healthy sample decided badly, poor perceivers still rated arousal and some still chose well, and the authors conclude that bodily feedback “may not be essential” and that people with poor interoception “make more use of alternative, nonbodily, mechanisms.”
- For this wiki specifically: the framework’s central assumption — that interoception can be left as an assumption (skin-conductance-response, somatic-marker-hypothesis) — is wrong on its own terms. Measure the perceiving and the somatic signal’s effect on behaviour changes sign. A theory of bodily signalling that never asks who can feel the signal has been leaving its largest moderator in the error term.
The authors’ closing proposal is the one the wiki should carry forward: stop asking whether cognitive-affective processes are embodied and start asking “when, to what extent, and for whom.” That is a research programme, and — see the limitations — this paper is one moderately-powered, unreplicated instalment of it.